Efficient Generation of Pretraining Samples for Developing a Deep Learning Brain Injury Model via Transfer Learning.

Area of Science:

• Computational neuroscience
• Machine learning in biomechanics
• Deep learning for injury modeling

Background:

• Deep learning models for brain injury analysis require substantial computational resources due to large training datasets.
• High accuracy in brain injury prediction is crucial for developing effective safety measures and protective equipment.

Purpose of the Study:

• To investigate the efficiency of using a transformer neural network (TNN) to generate pretraining samples for convolutional neural network (CNN) brain injury models.
• To reduce the computational cost associated with training deep learning brain injury models.

Main Methods:

• Validated a high-accuracy TNN for generating synthetic and augmented impact samples, emulating real-world and measured events.
• Utilized TNN-generated samples to pretrain a CNN, followed by finetuning with limited directly simulated data.
• Assessed CNN performance using an independent dataset of measured impact events.

Main Results:

• The TNN demonstrated high accuracy (0.948-0.967) in estimating voxelized peak strains for impact simulations.
• Pretraining with TNN-generated samples significantly improved CNN accuracy through transfer learning compared to baseline models.
• Optimal improvement was observed with 2000-4000 pretraining samples and a small finetuning dataset (500 samples), boosting success rate from 0.72 to 0.81.

Conclusions:

• The TNN can efficiently generate pretraining data, offering a cost-effective strategy for training deep learning brain injury models.
• This method facilitates wider adoption of deep learning for large-scale brain injury prediction, potentially enhancing safety protocols.
• The TNN-based pretraining approach reduces reliance on computationally expensive direct simulations.